Now you see it, now you don't: statistical and methodological considerations in fMRI.

We illustrate the effects of statistical threshold, spatial clustering, voxel size, and two approaches to multiple comparison correction on fMRI results. We first analyzed fMRI images obtained from a single subject during a noun-verb matching task. Data were analyzed with Statistical Parametric Mapping (SPM) using two different voxel sizes, and results were displayed at three different levels of statistical significance. At each statistical threshold, results were first uncorrected for multiple comparisons and spatial extent and then presented using a spatial extent cluster of 20 voxels. We then statistically controlled the Type I error rate associated with multiple comparisons by using the false discovery rate and by the random field adjustment for false-positive rate used by SPM. We also examined group results from language and graphesthesia paradigms at three levels of statistical significance. In all circumstances, apparent random activations decreased as more conservative statistical approaches were employed, but activation in areas considered to be functionally significant was also reduced. These issues are important in the choice of analytic approach and interpretation of fMRI results, with clear implications for the surgical management of individual patients when fMRI results are used to delineate specific areas of eloquent cortex.